A Scalable and Generalized Deep Ensemble Model for Road Anomaly Detection in Surveillance Videos

Surveillance cameras have been widely used for monitoring in both private and public sectors as a security measure. Close Circuits Television (CCTV) Cameras are used to surveillance and monitor the normal and anomalous incidents. Real-world anomaly detection is a significant challenge due to its complex and diverse nature. It is difficult to manually analyze because vast amounts of video data have been generated through surveillance systems, and the need for automated techniques has been raised to enhance detection accuracy. This paper proposes a novel deep-stacked ensemble model integrated with a data augmentation approach called Stack Ensemble Road Anomaly Detection (SERAD). SERAD is used to detect and classify the four most happening road anomalies, such as accidents, car fires, fighting, and snatching, through road surveillance videos with high accuracy. The SERAD adapted three pre-trained Convolutional Neural Networks (CNNs) models, namely VGG19, ResNet50 and InceptionV3. The stacking technique is employed to incorporate these three models, resulting in much-improved accuracy for classifying road abnormalities compared to individual models. Additionally, it presented a custom real-world Road Anomaly Dataset (RAD) comprising a comprehensive collection of road images and videos. The experimental results demonstrate the strength and reliability of the proposed SERAD model, achieving an impressive classification accuracy of 98.7%. The results indicate that the proposed SERAD model outperforms than the individual CNN base models.